Regression-Based Human Motion Capture From Voxel Data

Abstract

A regression based method is proposed to recover human body pose from 3D voxel data. In order to do this we need to convert the voxel data into a feature vector. This is done using a Bayesian approach based on Mixture of Probabilistic PCA that transforms a collection of 3D shape context descriptors, extracted from the voxels, to a compact feature vector. For the regression, the newly-proposed Multi-Variate Relevance Vector Machine is explored to learn a single mapping from this feature vector to a low-dimensional representation of full body pose. We demonstrate the effectiveness and robustness of our method with experiments on both synthetic data and real sequences.

Discriminative Human Full-Body Pose Estimation from Wearable Inertial Sensor Data

“…The quantitative evaluation of our method shows that the full-body pose can be estimated for the activities we considered (Section 3.1) with average errors of 4-6 degrees per joint, using four, three and even two inertial orientation sensors. This reconstruction quality is comparable to other state-of-the-art methods that use higher-dimesional input for discriminative pose estimation [12,13,16,17]. …”

“…Disadvantages of such generative pose estimation methods are their sensitivity to initialization and the high computational cost of pose inference [13]. Several authors therefore propose to directly learn a mapping from simple, lowdimensional observations to the corresponding full body poses by non-linear regression [12,17,15]. Such discriminative approaches have the advantage that poses can be predicted efficiently, without the need for iterative optimization.…”

“…There is a large variety of learning-based approaches for human pose estimation in the literature [12,13,14,15,16,17]. Many recently published methods are generative and focus on learning low-dimensional representations of human poses using so-called latent variable models [8,9,10,11].…”

“…Different types of input data have been proposed for discriminative pose estimation, such as silhouettes in 2D images [12], SIFT-features [15] or 3D voxelrepresentations of humans [17]. Fossati et al [13] use trajectories of body parts in videos as an input.…”

Real-Time Human Pose Tracking from Range Data

“…Learning-based approaches [6][7][8] require a large amount of training data to sample the large space of body configurations. Many papers have proposed to detect body parts from images as a preprocessing step for full-body pose reconstruction.…”